DDS Analytics

Talent Management Solutions

Frito Lay Attrition, Salary, and Job Roles

Executive Summary

EDA conclusions: Overtime, monthly income, stock option, job involvement, and years in current role were most useful for predicting attrition

Two Naive Bayes attrition prediction models (different options depending on intended use)

One regression monthly income prediction model using job level, role, and total working years

Additional observations regarding sales representatives, stock options, and high potential employees

Presentation

EDA

Visualization of Factors Associated with Attrition

## 
##  Pearson's Chi-squared test with Yates' continuity correction
## 
## data:  table(attrition$OverTime, attrition$Attrition)
## X-squared = 62.762, df = 1, p-value = 2.333e-15

##      
##             No      Yes
##   No  518.5517 99.44828
##   Yes 211.4483 40.55172

##      
##        No Yes
##   No  558  60
##   Yes 172  80

## 
##  Two Sample t-test
## 
## data:  MonthlyIncome by Attrition
## t = 4.6199, df = 868, p-value = 4.422e-06
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
##  1114.207 2760.222
## sample estimates:
##  mean in group No mean in group Yes 
##          6702.000          4764.786

## 
##  Two Sample t-test
## 
## data:  YearsInCurrentRole by Attrition
## t = 4.6596, df = 868, p-value = 3.665e-06
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
##  0.8949634 2.1976002
## sample estimates:
##  mean in group No mean in group Yes 
##          4.453425          2.907143

Employees who leave are more likely to:

(Top three factors)

Work overtime (X(1) = 62.76, p<.0001)

Have lower monthly income (t = 4.62, p<.0001)

Have lower job involvement scores

Also:

Have spent less time in their current role (t = 4.45, p<.0001)

Be in a Sales Rep, HR, or Lab Tech job role

Additional promising variables with respect to attrition:

Employees who leave are more likely to:

Be Single

Be Younger

Report worse (1) work life balance

Have less years at the company

Have less time with their current manager

Take stock options 0 and 3

Have less total working years

Have the lowest job level

Be further from home

Be less satisfied with their environment

Variables with less difference in attrition rate:

Employees who separate are more likely:

To have an education in HR, Marketing, or a Technical degree

Report less job satisfaction

Have more business travel

Work in the sales department

Variables without much difference in attrition rate:

Variables that do not appear helpful for predicting attrition:

Hourly Rate, Daily Rate, Monthly Rate, Percent Salary Hike, Performance Rating

Gender, Employee Number, Standard Hours, Over 18, Employee Count, ID

Examine correlations among quantitative variables:

Building a Model to Predict Attrition

Exclude:

Hourly Rate, Daily Rate, Monthly Rate, Percent Salary Hike, Performance Rating, Gender, Employee Number, Standard Hours, Over 18, Employee Count, and ID, as they have little or no difference in attrition

Reduce number of predictors:

From the EDA, Education, Education Field, Years Since Last Promotion, Relationship Satisfaction, Job Satisfaction, and Department appear to have less potential for predicting attrition.

Feature Selection Algorithm

## Boruta performed 99 iterations in 19.53276 secs.
##  15 attributes confirmed important: Age, Department, JobInvolvement,
## JobLevel, JobRole and 10 more;
##  19 attributes confirmed unimportant: BusinessTravel, DailyRate,
## DistanceFromHome, Education, EducationField and 14 more;
##  1 tentative attributes left: EnvironmentSatisfaction;

A random forest feature selection algorithm (Boruta) was then used to suggest additional variables to exclude from the final model.

At this point, I used trial and error to decide on the final predictors to include in the models.

Possible NB Models for Attrition

## [1] 0.7858621

## [1] 0.8237927

## [1] 0.5898709

## [1] 0.6915727

## [1] 0.6861773

## [1] 0.7730994

Two models to predict attrition:

Naive Bayes model with 13 variables:

Overtime, Monthly Income, Stock Option Level, Job Involvement, Years in Current Role, Total Working Years, Age, Job Level, Years at Company, Job Role, Years with Current Manager, Work Life Balance, Number Companies Worked

Mean sensitivity: 82%

Mean specificity: 60%

Useful if misidentifying retained employees has a high cost

Naive Bayes model with 5 variables:

As the data set was unbalanced (there was 16% attrition), this model was developed with a downsampling technique that balanced the number of retained and separated employees. It was then tested on an unbalanced test set to determine how well it will work on real world data that will most likely have more retained than separated employees.

Variables: Overtime, Monthly Income, Stock Option Level, Job Involvement, Years in Current Role

Mean sensitivity: 68%

Mean specificity: 77%

If better prediction of attrition is desirable, and the costs associated with incorrectly identifying retained employees is low, this model is useful.

Monthly Income and Job Level, Job Role, and Total Working Years

A visualization:

Previous correlation plot showed monthly income strongly correlated with:

Job Level, followed by Total Working Years

These two variables show a linear relationship with monthly income.

The addition of Job Role looks like it may improve the prediction, as different jobs have different monthly income ranges.

Regression Model to Predict Monthly Income

## 
## Call:
## lm(formula = MonthlyIncome ~ JobLevel + JobRole + TotalWorkingYears, 
##     data = attrition)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -3989.9  -631.3    -9.4   616.8  4144.6 
## 
## Coefficients:
##                               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)                   -103.243    203.291  -0.508  0.61168    
## JobLevel                      2795.329     81.784  34.179  < 2e-16 ***
## JobRoleHuman Resources        -321.962    251.744  -1.279  0.20127    
## JobRoleLaboratory Technician  -597.329    168.984  -3.535  0.00043 ***
## JobRoleManager                4002.415    228.696  17.501  < 2e-16 ***
## JobRoleManufacturing Director  137.309    167.212   0.821  0.41178    
## JobRoleResearch Director      3993.674    214.342  18.632  < 2e-16 ***
## JobRoleResearch Scientist     -324.522    169.387  -1.916  0.05572 .  
## JobRoleSales Executive         -66.048    144.221  -0.458  0.64709    
## JobRoleSales Representative   -410.303    211.660  -1.939  0.05289 .  
## TotalWorkingYears               47.998      7.951   6.037 2.34e-09 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 1063 on 859 degrees of freedom
## Multiple R-squared:  0.9472, Adjusted R-squared:  0.9466 
## F-statistic:  1541 on 10 and 859 DF,  p-value: < 2.2e-16

## [1] 1070.13

The combination of variables in this model explains 95% of the variance in monthly income.

Mean RMSE over 1000 tries: 1067

Additional observations regarding Job Roles, Over Time, and high potential employees

All single employees are in stock option 0. If this is voluntary, it is worth further examination. If it is a single employee’s only option, is this a policy or an oversight?

Sales Representatives have the shortest mean years at the company (2.9) with 50% between 1-3 years.

There is a strong association between total working years and job level. It appears as though there are a few long-term level 3 employees. There are also a few level 4 employees who have less than 20 working years. This may indicate they are high potential employees.

DDS Analytics

Nicole Norelli

11/24/2020

DDS Analytics

Talent Management Solutions

Frito Lay Attrition, Salary, and Job Roles

Executive Summary

EDA conclusions: Overtime, monthly income, stock option, job involvement, and years in current role were most useful for predicting attrition

Two Naive Bayes attrition prediction models (different options depending on intended use)

One regression monthly income prediction model using job level, role, and total working years

Additional observations regarding sales representatives, stock options, and high potential employees

EDA

Visualization of Factors Associated with Attrition

Employees who leave are more likely to:

(Top three factors)

Work overtime (X(1) = 62.76, p<.0001)

Have lower monthly income (t = 4.62, p<.0001)

Have lower job involvement scores

Also:

Have spent less time in their current role (t = 4.45, p<.0001)

Be in a Sales Rep, HR, or Lab Tech job role

Additional promising variables with respect to attrition:

Employees who leave are more likely to:

Be Single

Be Younger

Report worse (1) work life balance

Have less years at the company

Have less time with their current manager

Take stock options 0 and 3

Have less total working years

Have the lowest job level

Be further from home

Be less satisfied with their environment

Variables with less difference in attrition rate:

Employees who separate are more likely:

To have an education in HR, Marketing, or a Technical degree

Report less job satisfaction

Have more business travel

Work in the sales department

Variables without much difference in attrition rate:

Variables that do not appear helpful for predicting attrition:

Hourly Rate, Daily Rate, Monthly Rate, Percent Salary Hike, Performance Rating

Gender, Employee Number, Standard Hours, Over 18, Employee Count, ID

Examine correlations among quantitative variables:

Building a Model to Predict Attrition

Exclude:

Hourly Rate, Daily Rate, Monthly Rate, Percent Salary Hike, Performance Rating, Gender, Employee Number, Standard Hours, Over 18, Employee Count, and ID, as they have little or no difference in attrition

Reduce number of predictors:

From the EDA, Education, Education Field, Years Since Last Promotion, Relationship Satisfaction, Job Satisfaction, and Department appear to have less potential for predicting attrition.

Although years with current manager, in current role, and at company all share noticable correlation with each other, removing any of them worsened the model predictions.

Feature Selection Algorithm

A random forest feature selection algorithm (Boruta) was then used to suggest additional variables to exclude from the final model.

At this point, I used trial and error to decide on the final predictors to include in the models.

Possible NB Models for Attrition

Two models to predict attrition:

Naive Bayes model with 13 variables:

Overtime, Monthly Income, Stock Option Level, Job Involvement, Years in Current Role, Total Working Years, Age, Job Level, Years at Company, Job Role, Years with Current Manager, Work Life Balance, Number Companies Worked

Mean sensitivity: 82%

Mean specificity: 60%

Useful if misidentifying retained employees has a high cost

Naive Bayes model with 5 variables:

Variables: Overtime, Monthly Income, Stock Option Level, Job Involvement, Years in Current Role

Mean sensitivity: 68%

Mean specificity: 77%

If better prediction of attrition is desirable, and the costs associated with incorrectly identifying retained employees is low, this model is useful.

Monthly Income and Job Level, Job Role, and Total Working Years

A visualization:

Previous correlation plot showed monthly income strongly correlated with:

Job Level, followed by Total Working Years

These two variables show a linear relationship with monthly income.

The addition of Job Role looks like it may improve the prediction, as different jobs have different monthly income ranges.

Regression Model to Predict Monthly Income

The combination of variables in this model explains 95% of the variance in monthly income.

Mean RMSE over 1000 tries: 1067

Additional observations regarding Job Roles, Over Time, and high potential employees

All single employees are in stock option 0. If this is voluntary, it is worth further examination. If it is a single employee’s only option, is this a policy or an oversight?

Sales Representatives have the shortest mean years at the company (2.9) with 50% between 1-3 years.

There is a strong association between total working years and job level. It appears as though there are a few long-term level 3 employees. There are also a few level 4 employees who have less than 20 working years. This may indicate they are high potential employees.